Efficiently Fit Generalized Linear Mixed Model with a Single Random Effect on Massive Datasets

Fitting linear mixed models (LMMs) and generalized linear mixed models (GLMMs) on large-scale datasets can be very computationally expensive. The fastglmm package provides a very fast method for fitting LMMs and GLMMs with a single random effect. This can be 10-1000x faster than other methods for large datasets.

The fastglmm package is designed to be intuitive for users of other mixed model software such as lme4, and uses familar mixed model syntax and R generics. Since fastglmm is designed for large datasets and GLMMs in mind, it uses maximum likelihood rather than REML to fit LMMs.

Example Usage

As an introduction, consider fitting a LMM to the sleepstudy data distributed with lme4. The model syntax and parameter estimates are the same as using lme4::lmer(..., REML=FALSE).

library(fastglmm)
data(sleepstudy, package="lme4")

fit <- fastlmm(Reaction ~ Days + (1 | Subject), sleepstudy)

fit
#> 
#> Call:
#> fastlmm(formula = Reaction ~ Days + (1 | Subject), data = sleepstudy)
#> 
#> Coefficients:
#> (Intercept)         Days  
#>      251.41        10.47

summary(fit)
#> Linear mixed model fit by ML ['fastlmm']
#>  Formula: Reaction ~ Days + (1 | Subject)
#> 
#> Coefficients:
#>             Estimate Std. Error   ddf t value Pr(>|t|)    
#> (Intercept) 251.4051     9.5062  24.5   26.45   <2e-16 ***
#> Days         10.4673     0.8017 162.0   13.06   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Residual df: 162.2 
#> 
#> Variance components:
#>   sigSq_g: 1297
#>   sigSq_e: 954.5
#>   delta:   0.736

Installation

devtools::install_github("GabrielHoffman/fastglmm")

Implementation details

Analysis of a given response or features is performed on a single thread. Analysis is parallelized across responses/features using oneAPI Threading Building Blocks (oneTBB)
Linear algebra is performed with Armadillo
Use of dense and sparse matricies is supported at the C++ level using templates
Interface between R and C++ is implemented in Rcpp and also depends on RcppArmadillo and RcppParallel
Implemented in C++17